library(OpenImageR)
library(mxnet)
library(magick)
library(stringi)set.seed(131242351)
modify_image <- function(flip_mode, shift_cols, shift_rows, rotate_angle, new_image_name, patch) {
if (is.na(flip_mode)) {
patchAugmented <- Augmentation(patch, shift_cols = shift_cols, shift_rows = shift_rows,
rotate_angle = rotate_angle, rotate_method = 'bilinear',
zca_comps = 30,zca_epsilon = 0.1, threads = 1, verbose = F)
} else {
patchAugmented <- Augmentation(patch, flip_mode = flip_mode,
shift_cols = shift_cols, shift_rows = shift_rows,
rotate_angle = rotate_angle, rotate_method = 'bilinear',
zca_comps = 30,zca_epsilon = 0.1, threads = 1, verbose = F)
}
writeImage(patchAugmented, new_image_name)
}
for (file in list.files("patches", full.names = T)) {
print(file)
patch <- readImage(file)
file_basename <- stri_split_fixed(basename(file), '.')[[1]][1]
file_dir <- dirname(file)
flip_mode <- sample(c('horizontal', 'vertical', NA), size = 50, replace = T)
shift_cols <- sample(-15:15, size = 50, replace = T)
shift_rows <- sample(-15:15, size = 50, replace = T)
rotate_angle <- sample(seq(0, 360, by = 15), size = 50, replace = T)
new_file_names <- c()
for (i in 1:50) {
new_file_names[i] <- stri_join("modified_patches/", file_basename, "_", i,".jpg")
}
mapply(modify_image, flip_mode, shift_cols, shift_rows, rotate_angle, new_file_names, MoreArgs = list(patch))
print(file)
}data <- mx.symbol.Variable('data')
conv.0 <- mx.symbol.Convolution(data = data, kernel = c(5, 5), num_filter = 10)
tanh.0 <- mx.symbol.Activation(conv.0, act_type="tanh")
pool.0 <- mx.symbol.Pooling(data=tanh.0, kernel=c(2, 2), stride=c(2, 2), pool.type="max")
conv.1 <- mx.symbol.Convolution(data = pool.0, kernel = c(5, 5), num_filter = 10)
tanh.1 <- mx.symbol.Activation(conv.1, act_type="tanh")
pool.1 <- mx.symbol.Pooling(data=tanh.1, kernel=c(2, 2), stride=c(2, 2), pool.type="max")
fc.0 <- mx.symbol.FullyConnected(data = pool.1, num_hidden =3)
nn.model <- mx.symbol.SoftmaxOutput(data = fc.0)
graph.viz(nn.model)# Select train samples from initial 168 files
training <- sample(1:168, 134, replace = F)
training_full <- rep(0, 134*50)
# Since we've modified initial files 50 times, train samples should be corrected
for (i in 1:134) {
training_full[((50*(i - 1)) + 1):(i * 50)] <- (50 * (training[i] - 1) + 1 ):(50 * training[i])
}
testing <- (1:(168*50))[-training_full]
features <- 61 * 61
dataset.size <- 168 * 50
nn.data.x <- matrix(0, nrow=dataset.size, ncol=features)
nn.data.y <- vector(length=dataset.size)
# Read data
for (i in 1:168) {
for (j in 1:50) {
print(sprintf("modified_patches/patch%s_%s.jpg", i, j))
nn.data.x[(i - 1) * 50 + j, ] <- as.numeric(readImage(sprintf("modified_patches/patch%s_%s.jpg", i, j)))
}
}
nn.data.y <- rep(read.csv("patch_labels.csv", header = F)$V1, each = 50)
train.x <- nn.data.x[training_full, ]
train.y <- nn.data.y[training_full]
test.x <- nn.data.x[testing, ]
test.y <- nn.data.y[testing]
train.array <- t(train.x)
dim(train.array) <- c(61, 61, 1, ncol(train.array))
test.array <- t(test.x)
dim(test.array) <- c(61, 61, 1, ncol(test.array))mx.set.seed(3774)
model <- mx.model.FeedForward.create(nn.model,
X = new.train.array,
y = as.array(train.y - 1),
eval.data = list(
data = new.test.array,
label = as.array(test.y - 1)#test.y
),
ctx = mx.cpu(),
num.round = 200,
optimizer = "adadelta",
eval.metric = mx.metric.accuracy,
epoch.end.callback = mx.callback.log.train.metric(10))Start training with 1 devices
[1] Train-accuracy=0.348106971153846
[1] Validation-accuracy=0.236049107142857
[2] Train-accuracy=0.34094929245283
…
[70] Train-accuracy=0.531544811320755
[70] Validation-accuracy=0.309151785714286
[71] Train-accuracy=0.498378537735849
[71] Validation-accuracy=0.335379464285714
[72] Train-accuracy=0.481426886792453
[72] Validation-accuracy=0.440290178571429
[73] Train-accuracy=0.446639150943396
[73] Validation-accuracy=0.329799107142857
[74] Train-accuracy=0.520931603773585
[74] Validation-accuracy=0.41796875
[75] Train-accuracy=0.519162735849057
[75] Validation-accuracy=0.328683035714286
[76] Train-accuracy=0.516214622641509
[76] Validation-accuracy=0.299665178571429
[77] Train-accuracy=0.448260613207547
[110] Train-accuracy=0.626031839622642
[110] Validation-accuracy=0.339285714285714
[111] Train-accuracy=0.609817216981132
[111] Validation-accuracy=0.295200892857143
[112] Train-accuracy=0.6015625
[112] Validation-accuracy=0.342075892857143
[113] Train-accuracy=0.40123820754717
[113] Validation-accuracy=0.41015625
[114] Train-accuracy=0.518278301886792
[114] Validation-accuracy=0.329241071428571
[115] Train-accuracy=0.645931603773585
[115] Validation-accuracy=0.340401785714286
[116] Train-accuracy=0.638119103773585
[116] Validation-accuracy=0.337053571428571
[117] Train-accuracy=0.619103773584906
[117] Validation-accuracy=0.248883928571429
[118] Train-accuracy=0.448408018867925
[118] Validation-accuracy=0.350446428571429
[119] Train-accuracy=0.620725235849057
[119] Validation-accuracy=0.337611607142857
[120] Train-accuracy=0.645931603773585
[120] Validation-accuracy=0.348772321428571
[121] Train-accuracy=0.645784198113208
[121] Validation-accuracy=0.327566964285714
[122] Train-accuracy=0.636645047169811
[122] Validation-accuracy=0.329241071428571
[123] Train-accuracy=0.623083726415094
[123] Validation-accuracy=0.306361607142857
[124] Train-accuracy=0.660524764150943
[124] Validation-accuracy=0.345424107142857
[125] Train-accuracy=0.664652122641509
[125] Validation-accuracy=0.322544642857143
[126] Train-accuracy=0.668042452830189
[126] Validation-accuracy=0.328125
[127] Train-accuracy=0.658903301886792
[127] Validation-accuracy=0.310267857142857
…
[200] Train-accuracy=0.78626179245283
[200] Validation-accuracy=0.315848214285714